sathiskumarjothi0/Beveridge-Pro-C

A System One decision model with open weights. Give it a document (text, email, ticket, log or JSON) and typed questions; it returns typed answers (yes/no, a choice between options you describe, or a score on levels you define) with a calibrated probability for every option: one scoring pass per question, with no reasoning steps. It never generates text, so there is nothing to parse and nothing to hallucinate. It runs on your own machines: your documents never leave them.

About 26B parameters, built on Google Gemma 4 26B-A4B (mixture of experts, about 4B active per token). Everything it needs is in this repository, so nothing else is downloaded. English text, up to 4,096 input tokens; needs one GPU with about 60 GB of memory (bf16).

Model	Size	Runs on	Best for	Status
Beveridge-Large-C	0.4B (ModernBERT-large base)	laptop CPU/GPU, edge boxes, any GPU	operations documents, logs and alerts, counting, amounts and limits, ticket routing, request safety	available
Beveridge-Pro-C	26B (Gemma 4 26B-A4B base)	one GPU with about 60 GB	knowledge-heavy decisions: contracts, finance topics, large intent sets, dates and payment status	available

Installation

python -m pip install "bev-decisions[pro]"

Python 3.10 or newer. Optional extras: bev-decisions[serve] (HTTP server), bev-decisions[mcp] (MCP server for AI assistants and agents), bev-decisions[structured] (Pydantic schemas), bev-decisions[all]. The first bev.load() downloads the model from this repository and caches it; after that it works offline.

Quickstart

import bev

model = bev.load("sathiskumarjothi0/Beveridge-Pro-C")
answers = model.decide(
    "Subject: Charged twice for order 88412. I was billed $49.99 twice on 3 October. Please refund one payment today.",
    {
        "team":     {"type": "choice", "instructions": "Which team should handle this ticket?",
                     "criteria": {"billing": "payments, invoices, refunds", "tech": "bugs, outages, errors", "sales": "pricing and plans"}},
        "urgent":   {"type": "noul", "instructions": "Does the customer ask for urgent handling?"},
        "severity": {"type": "score", "instructions": "How severe is the problem for the customer?", "criteria": ["low", "medium", "high"]},
    },
)
answers["team"]   # {"choice": ..., "confidence": ..., "probabilities": {...}}

Run with or without a GPU

Beveridge-Pro-C needs a GPU: about 60 GB of GPU memory (for example one A100 80 GB or H100). bev.load() uses it automatically, or set it yourself:

model = bev.load("sathiskumarjothi0/Beveridge-Pro-C", device="cuda")   # NVIDIA GPU with about 60 GB

Measured on one NVIDIA H200 (package from PyPI): about 0.07 s per document with 3 questions (22 ms per decision; 7.8 ms per decision in batches of 64). The first load reads about 52 GB of weights.

Without a large GPU, use Beveridge-Large-C instead: it runs on a laptop CPU (about 0.18 s per document with 3 questions on an Apple M4 Pro) or GPU. Route mode (bev.Router(load_strong=True)) runs Beveridge-Large-C first and sends only the unsure answers to Beveridge-Pro-C on the GPU machine.

Use cases

Use case	Example questions	Where the evidence is
Support inbox triage	which team (choice), urgent? (yes/no), customer sentiment (score)	Beveridge-Bench ticket questions; support intents
Request safety and guardrails	is this request harmful? is it a jailbreak attempt?	Aegis 2.0 and JailbreakBench
AI agents: tool dispatch and checks	does this need a tool? which tool? is this action risky? (MCP server: bev mcp)	typed decisions
Model routing	which handler: rules, a small model, a large LLM or a person? how hard is this request?	typed decisions
Security and operations logs	attack category, failed logins above a number, severity, outage root cause	Beveridge-Bench security and alert questions
Back office documents	invoice total over the approval limit? how many items above a price?	Beveridge-Bench invoice questions
Contract and policy review	does this clause look unfair to the customer? which policy rule applies?	unfair contract terms; Rules-Bench
Reviews and content moderation	review sentiment, toxic comment?	product reviews; toxic comments

Check any use case on a few hundred labelled examples of your own first (bev eval my_cases.jsonl).

Type	Use it for	Answer
noul	yes or no	noul: probability of yes
choice	one of the options you define (criteria: key → description, up to 300)	choice, confidence, probabilities
score	ordered levels (criteria: list or key → description, lowest first)	label, score (expected level), confidence, probabilities

Up to 16 questions per call; Beveridge-Pro-C reads the document once per question. Questions with more than 26 options take a second round (a winner per group of 26, then a final).

Confidence gating

sure, review = bev.gate(answers, threshold=0.8)   # act on `sure`, send `review` to a person

Route mode

Route mode sends every question to Beveridge-Large-C first and only the unsure answers to Beveridge-Pro-C:

router = bev.Router(load_strong=True, threshold=0.8)
result = router.decide(text, questions)
result["answers"], result["routing"]          # routing: {"team": "fast", "urgent": "strong"}

Schema-driven decisions

from typing import Literal
from pydantic import BaseModel, Field

class Ticket(BaseModel):
    urgent: bool = Field(description="Does the customer ask for urgent handling?")
    team: Literal["billing", "tech", "sales"] = Field(description="Which team should handle this ticket?")

ticket = bev.decide_as(model, text, Ticket)       # Ticket(urgent=True, team='billing')

Self-hosting: HTTP and MCP

pip install "bev-decisions[serve]"
bev serve --model sathiskumarjothi0/Beveridge-Pro-C --port 8000      # POST /v1/decide, /v1/batch; GET /v1/presets, /health
pip install "bev-decisions[mcp]"
bev mcp --model sathiskumarjothi0/Beveridge-Pro-C                    # tools: decide, decide_preset, list_presets

The server listens on 127.0.0.1 without authentication; put it behind your own gateway before exposing it. Docker files and examples are in the GitHub repository.

Benchmarks

Accuracy (%) = correct answers ÷ questions on held-out test sets; the same cases for every system (October 2026). Bold = best in the row; – = not run. * = an area Beveridge was not built for.

Decisions on documents

Test	Beveridge-Large-C	Beveridge-Pro-C	Jev	Mistral Large 4	Mistral Large 3	gpt-oss 120B	Gemma 4 31B	Nemotron 3 Super
Beveridge-Bench (operations documents)	93.0	91.9	87.9	83.7	74.6	91.1	88.9	70.6
Typed decisions	75.8	79.2	73.8	72.0	68.0	68.8	75.0	62.7
Rules-Bench (rules shown)	85.6	92.0	92.3	83.1	81.4	94.5	91.6	82.7
JailbreakBench refusal	84.5	84.0	83.5	77.5	62.0	68.0	87.5	65.0
Aegis 2.0 request safety	82.3	78.2	71.5	70.5	66.9	69.5	73.1	69.3

Public classification tests

Test	Beveridge-Large-C	Beveridge-Pro-C	Jev	Mistral Large 4	Mistral Large 3	gpt-oss 120B	Gemma 4 31B	Nemotron 3 Super
AG News	86.6	90.6	91.8	90.8	88.8	88.8	91.8	87.8
Emotion	53.2	56.2	57.6	55.8	55.6	57.8	55.4	50.2
Banking77 (77 intents)	63.6	70.6	79.2	75.8	71.2	81.0	78.8	69.2
SST-5 (5-level sentiment)	33.2	51.8	56.8	51.4	49.4	49.6	53.2	46.8
XNLI (English)	84.6	86.4	89.2	84.8	77.6	80.2	89.2	75.0
MASSIVE intents	59.4	74.8	80.8	78.4	75.2	81.6	83.8	69.4

Areas Beveridge was not built for

Test	Beveridge-Large-C	Beveridge-Pro-C	Jev	Mistral Large 4	Mistral Large 3	gpt-oss 120B	Gemma 4 31B	Nemotron 3 Super
Product reviews*	93.2	96.4	96.4	96.4	96.6	95.4	97.2	94.8
Unfair contract terms*	65.6	76.0	62.0	54.6	52.8	47.6	74.4	44.0
Financial-news sentiment*	69.8	59.2	66.4	64.0	50.4	68.4	71.2	64.0
Poem sentiment*	71.8	65.1	44.5	42.6	32.1	59.8	62.7	45.5
Support intents*	85.6	88.0	89.8	85.4	86.2	90.0	90.4	76.6
Toxic comments*	63.0	79.2	78.8	68.0	73.0	80.8	76.4	69.6
Financial-news topics*	37.8	69.8	65.0	61.8	61.6	63.4	64.2	57.2
Legal case holdings*	42.0	71.2	77.6	69.2	72.2	61.2	71.6	55.0

Size and speed

Accuracy on the same 10,238 answers against total parameters (log scale), for every system with a published size: Beveridge-Large-C (0.4B) and Beveridge-Pro-C (25.8B) are more accurate than LLMs of 31B to 675B.

Beveridge, one document with 3 questions (package from PyPI)	Per decision	In batches of 64
Beveridge-Large-C, one NVIDIA H200 GPU	4.9 ms	3.4 ms
Beveridge-Large-C, Apple M4 Pro laptop GPU	17 ms
Beveridge-Large-C, Apple M4 Pro laptop CPU	60 ms
Beveridge-Pro-C, one NVIDIA H200 GPU	22 ms	7.8 ms

The LLMs and Jev were measured through their cloud APIs (median response time per decision, including network and queueing): Jev 98 ms, the five LLMs 4.1 to 7.0 s.

How much of the work can you automate?

Sort every answer from most to least confident and hand over the most confident ones first. The share you can hand over while those decisions stay at least 95% (or 90%) correct is the automation rate. Example: "76% at 95%" means 3 of every 4 decisions are taken automatically and at least 95% of them are right; the rest go to a person. Answers given with the same confidence are handed over together, so a system whose "100% sure" answers are only 94% right reaches 0% at the 95% target. Held-out tests: Beveridge-Bench, typed decisions and Aegis 2.0 request safety, 10,238 answers per system.

System	Automated at 95% accuracy	Automated at 90% accuracy	Accuracy on all answers
Beveridge-Large-C	76%	93%	87.0%
Beveridge-Pro-C	69%	90%	86.0%
Jev	59%	77%	81.1%
gpt-oss 120B	0%	74%	81.4%
Gemma 4 31B	0%	71%	82.2%
Mistral Large 4	0%	0%	78.1%
Mistral Large 3	0%	0%	71.4%
Nemotron 3 Super	0%	0%	68.7%

Beveridge-Bench, question by question

Averages hide weak spots. Bold = within 10 points of always answering the most common label, i.e. little or no skill.

Question (300 cases each)	Beveridge-Pro-C	Always-most-common baseline
Invoice: payment status vs due date	100.0%	38.3%
Security log: attack category	100.0%	20.7%
Security log: severity	100.0%	39.0%
Outage alert: root cause	100.0%	5.3%
Outage alert: SLO breached?	100.0%	82.0%
Ticket: urgent?	100.0%	66.0%
Which date is earlier	100.0%	50.0%
Choosing among many options	100.0%	0.7%
Comparing two amounts	99.3%	52.0%
Ticket: which department	94.7%	16.0%
Security log: failed-login count	91.3%	59.7%
Invoice: how late	89.3%	35.3%
Invoice: total over a limit?	85.0%	51.3%
Invoice: items above a price	82.7%	14.7%
Long log: intrusion?	74.0%	52.7%
Ticket: sentiment	70.3%	24.3%
Counting items in a list	62.7%	10.0%

Calibration

Test	Accuracy	Answered at confidence ≥ 0.8	Accuracy of those	ECE
Beveridge-Bench (operations documents)	91.9%	89.7%	95.0%	0.032
Rules-Bench (rules shown)	92.0%	87.2%	97.0%	0.011
Aegis 2.0 request safety	78.2%	51.1%	92.6%	0.024
Product reviews	96.4%	93.0%	97.9%	0.014
Support intents	88.0%	98.8%	88.3%	0.105
Typed decisions	79.2%	61.7%	91.5%	0.040

ECE (expected calibration error) = Σ over 10 confidence bins of (share of answers in the bin × |accuracy − mean confidence|); 0 is perfect. Example: a bin holding 40% of the answers at mean confidence 0.95 with 90% of them right adds 0.4 × 0.05 = 0.02. Calibration temperatures: noul 1.0, choice 1.2, score 1.2.

Package check

This repository, loaded with the bev-decisions package, was checked against the official evaluation run (accuracy = correct ÷ questions on the same cases):

Test	Questions	Package	Official run
Beveridge-Bench (operations documents)	1,200	88.9%	89.2%
Typed decisions	1,500	79.9%	79.3%
Product reviews	300	97.3%	98.0%

A few answers that are close to a tie can change with batch size on a mixture-of-experts model in bf16; accuracy stays within 1 point.

Honest limits

• Knowledge-heavy decisions are not solved: legal case holdings 71.2%, financial-news topics 69.8%.
• Questions with more than 26 options are answered as a tournament (a winner per group of 26, then a final); accuracy is lower there (Banking77 70.6%) and confidence less well calibrated.
• Five-level sentiment is hard (SST-5 51.8%).
• Needs a large GPU; for laptops and edge boxes use Beveridge-Large-C.
• English only; text beyond 4,096 tokens is cut; no images.
• Not for medical, legal or financial advice. For decisions with serious consequences, validate on your own data (bev eval my_cases.jsonl) and keep a person in the loop.

Files

File	Contents
model-*.safetensors, config.json, tokenizer files	Gemma 4 26B-A4B, unchanged (bf16)
adapter/	Beveridge's decision weights, loaded together with Gemma's
bev_config.json	Beveridge settings and calibration temperatures
NOTICE	Gemma attribution (Apache-2.0)
assets/	the logo and the charts on this page

Links

• Package: pypi.org/project/bev-decisions
• Code and examples: github.com/sathiskumarjothi0-oss/bev-decisions-pypi
• Beveridge-Large-C: huggingface.co/sathiskumarjothi0/Beveridge-Large-C

Support

If Beveridge is useful to you, you can support its development: buymeacoffee.com/sjothi

Licence and credit

Apache-2.0, © 2026 Dr. Sathiskumar JOTHI. Built on Gemma 4 26B-A4B (Apache-2.0) by Google.

About the name

Beveridge is named after Sir William Beveridge (1879–1963), the British economist, born in British India and educated in England, whose 1942 Beveridge Report laid the foundations of the modern welfare state and of the National Health Service (NHS), built for the good of future generations. The name is a reminder of what we want from the model: dependable, fair decisions that serve people, made carefully and checked against the facts.

Other models on Hugging Face are called Bev; they are not related to this model. The Python package keeps its short name: pip install bev-decisions, import bev.